Method for controlling a fuel cell system

Abstract

The invention relates to a method for controlling a fuel cell system (100). The fuel cell system (100) has a fuel cell stack (10) with an anode side (11) and a cathode side (12), an exhaust line (13), a fuel line with a recirculation circuit (14), and at least one valve line (15) connected to the recirculation circuit (14). The at least one valve line (15) and the exhaust line (13) merge into a measuring line (16). The at least one valve line (15) has a valve (18). The method comprises the following steps:-(S1) measuring an H2 concentration and / or a H2O concentration in the measuring line (16),-(S2) determining an H2 concentration, an N2 concentration, a steam concentration and / or an amount of water on the anode side (11) of the fuel cell stack (10) by means of a trained machine learning method (20) on the basis of the measured H2 concentration and / or H2O concentration, and-(S3) adjusting a purge duration and / or a purge interval on the basis of the determined concentration of H2 and / or N2 on the anode side (11) of the fuel cell stack (10) and adjusting a bleed duration and / or a bleed interval on the basis of the determined concentration of steam and / or water on the anode side (11) of the fuel cell stack (10), or-(S4) adjusting a bleed time and / or a bleed interval on the basis of the determined steam concentration and / or water amount on the anode side (11) of the fuel cell stack.

Description

Technical Field

[0001] The present invention generally relates to the field of fuel cells, in particular PEM fuel cells. In particular, the present invention relates to a method for controlling a fuel cell system, a method for training a machine learning method for a method for controlling a fuel cell system, and a fuel cell system. Background Art

[0002] In polymer electrolyte membrane fuel cells (PEM fuel cells) that operate with hydrogen and air, a proton-conducting membrane separates the gases of the two primary half-cells. The anode side is filled with hydrogen, and the cathode side is filled with air. Gas transfer occurs due to concentration differences and an incompletely gas-tight membrane.

[0003] The accumulation of nitrogen on the anode side leads in particular to a decrease in the hydrogen concentration there. This not only results in a reduced voltage and thus reduced efficiency, but also in the risk of hydrogen depletion, which can lead to irreversible damage.

[0004] According to the prior art, the nitrogen concentration on the anode side is kept low by regularly venting a portion of the anode gas through a valve and replacing it with pure hydrogen. This process is called "purging." The valve opening and closing times are based on empirical values ​​and are typically stored in the system's purge strategy as a function of current.

[0005] Furthermore, liquid water on the anode side must be regularly removed. This removal of liquid water can be referred to as a "bleeding" process. If this bleeding is performed too infrequently, it can lead to hydrogen depletion. Conversely, too frequent bleeding can lead to wasted hydrogen. Excessive purging or bleeding can reduce the vapor concentration on the anode side below a critical concentration, potentially causing the fuel cell to dry out.

[0006] By heuristically controlling the purge process and / or the bleed process, it is neither possible to ensure that the process is sufficiently far from the depletion boundary nor to ensure that the purge strategy or the bleed strategy is optimized in terms of hydrogen consumption. Summary of the Invention

[0007] The method according to the invention for controlling a fuel cell system having the features of the independent claim has the advantage that the purge strategy and the bleed strategy or the bleed strategy can be improved precisely, accurately and reliably with the aid of H 2 sensors already present for safety reasons.

[0008] The proposed method is less expensive, more reliable and more precise than prior art methods. The proposed method furthermore enables the provision of a fuel cell system which has a longer service life than prior art fuel cell systems.

[0009] The features and details described in connection with the method according to the invention for controlling a fuel cell system naturally also apply to the method according to the invention for training a machine learning method, the fuel cell system, the computer program product and / or the computer-readable medium, and vice versa, so that the disclosures concerning the various aspects of the invention are always mutually referenced or mutually referenceable.

[0010] A first aspect of the present invention relates to a method for controlling a fuel cell system, in particular a PEM fuel cell system. The fuel cell system comprises a fuel cell stack with an anode side and a cathode side, an exhaust gas line, a fuel line with a recirculation circuit, and at least one valve line connected to the recirculation circuit. The at least one valve line and the exhaust gas line merge into a measuring line. The at least one valve line comprises a valve. The method comprises the following steps:

[0011] - measuring the H2 concentration and / or H2O concentration in the measuring line,

[0012] - determining the H 2 concentration, N 2 concentration, steam concentration and / or water amount on the anode side of the fuel cell stack by means of a trained machine learning method based on the measured H 2 concentration and / or the measured H 2 O concentration, and

[0013] - adjusting the purge duration and / or purge interval based on the determined H2 concentration and / or N2 concentration on the anode side of the fuel cell stack and adjusting the bleed duration and / or bleed interval based on the determined steam concentration and / or water amount on the anode side of the fuel cell stack, or

[0014] - Adjusting the bleed duration and / or bleed interval based on the determined steam concentration and / or water amount on the anode side of the fuel cell stack.

[0015] In other words, the purge strategy and the bleed strategy or bleed strategy can be defined, optimized and / or adjusted in such a way that the purge duration and / or purge interval and the bleed duration and / or bleed interval or the optimized purge duration or bleed duration or the optimized purge interval or bleed interval can be determined. The purge strategy can define "how", "whether", "how often", "how long" and / or "when" the purge process is carried out. The bleed strategy can define "how", "whether", "how often", "how long" and / or "when" the bleed process is carried out.

[0016] Furthermore, a method can be provided that allows measurement data from existing sensors of a fuel cell system to be used in conjunction with a trained machine learning method to accurately estimate the hydrogen concentration on the anode side of the fuel cell system. Thus, a virtual measurement of the hydrogen concentration, nitrogen concentration, steam concentration, and / or water content on the anode side can be performed using a trained machine learning method. The trained machine learning method can be an artificial intelligence model.

[0017] The H2 concentration and / or H2O concentration in the measuring line can be measured using an H2 sensor. In particular, the H2 concentration and / or H2O concentration in the measuring line can be measured using existing sensors, as every fuel cell system is typically equipped with an H2 sensor in the exhaust line to measure or monitor the H2 concentration for safety reasons. Such an H2 sensor can also be configured to detect the vapor concentration in the measuring line. Furthermore, it is also conceivable to provide a sensor in the measuring line that is configured to detect the vapor concentration and / or the amount of water.

[0018] The measurement of H2 concentration and / or HO concentration can generate two signals, namely an H2 concentration signal and / or an HO concentration signal, and can be fed to the trained machine learning method. In other words, the measurement of H2 concentration and / or HO concentration can generate input signals for the trained machine learning method. Additionally, steam concentration can be measured. Such a measurement can generate a corresponding signal, namely a steam concentration signal, and feed it to the trained machine learning method. However, such a steam concentration signal is merely optional.

[0019] The trained machine learning method can determine the H2 concentration, N2 concentration, steam concentration and / or water amount on the anode side of the fuel cell stack. The concept of "determine" is to be understood broadly in the context of the present invention. Determining the H2 concentration, N2 concentration, steam concentration and / or water amount on the anode side can be a very precise and accurate assessment of the H2 concentration, N2 concentration, steam concentration and / or water amount on the anode side. Accordingly, the trained machine learning method can output a signal containing information about the H2 concentration, N2 concentration, steam concentration and / or water amount. Based on such a signal, the control unit of the fuel cell system can, for example, optimize the purge strategy and the bleed strategy or the bleed strategy by adjusting or determining the purge duration or the bleed duration and / or the purge interval or the bleed interval.

[0020] The expression "based on" is to be understood broadly in the context of the present invention, and may be understood to mean mutual correlation, dependency and / or ratio.

[0021] Such a method essentially provides a virtual mass spectrometer, since the data provided by a mass spectrometer that may be arranged in the recirculation loop can be determined using a trained machine learning method. Such a method essentially provides a virtual mass spectrometer, since the data provided by a water meter that may be arranged in the recirculation loop can or will be determined using a trained machine learning method.

[0022] Such a method thus enables feedback-regulated purge and bleed strategies or the determination of bleed strategies. Thus, the fuel cell system is protected from hydrogen depletion, allowing such a fuel cell system to enjoy a longer service life. Furthermore, hydrogen waste due to unnecessary purge and / or bleed processes can be reduced.

[0023] Advantageously, the valve of at least one valve line is a purge and drain valve. In other words, a single valve can be provided on the valve line, which is provided for not only performing a purge process but also a drain process.

[0024] Advantageously, the fuel cell system has two valve lines, wherein the first valve line has a purge valve and the second valve line has a drain valve. The first valve line can therefore be referred to as a purge line and the second valve line as a drain line.

[0025] Advantageously, the trained machine learning method is a Gaussian process model, in particular a Gaussian process model with a NARX structure.

[0026] Advantageously, the machine learning method is trained using a training fuel cell system. The training fuel cell system comprises a fuel cell stack having an anode side and a cathode side, an exhaust line, a fuel line having a recirculation circuit, at least one valve line connected to the recirculation circuit, a mass spectrometer, and a water meter. The mass spectrometer is configured to detect the H2 concentration and / or N2 concentration on the anode side of the training fuel cell system. The water meter is configured to detect the amount of water on the anode side of the training fuel cell system.

[0027] In other words, a training fuel cell system can be provided that can be used to train a trained machine learning method. The training fuel cell system can essentially differ from the fuel cell system in that it also has a mass spectrometer in the recirculation circuit. The training fuel cell system can also have a water meter in the recirculation circuit, which is configured to detect the water content, in particular the steam concentration. The training fuel cell system can have a bleed line, via which the water content can be detected (e.g., only on a test bench). Such a bleed line can be connected in parallel with the valve line.

[0028] Thus, a reliable, trained machine learning method can be provided that can be used to determine the H2 concentration, N2 concentration, steam concentration, and / or water content on the anode side. Thus, the H2 concentration, N2 concentration, steam concentration, and / or water content can be reliably and accurately determined again. This enables improved control of the fuel cell system. This also enables the provision of precise and reliable purge strategies and bleed strategies or bleed strategies.

[0029] Better control of the fuel cell system can lead to more precise and improved purge and / or bleeding strategies. This, in turn, can result in a longer service life for the corresponding fuel cell system. The purge and / or bleeding strategies can thus be optimized, resulting in less hydrogen being wasted during operation of the corresponding fuel cell system.

[0030] Advantageously, the method further comprises the following steps:

[0031] - deriving at least one characteristic of the respective measurement from the measured H 2 concentration and / or from the measured H 2 O concentration, and

[0032] - Based on the at least one feature derived from the corresponding measurement, determining the H 2 concentration, N 2 concentration, steam concentration and / or water amount on the anode side of the fuel cell stack by means of a trained machine learning method.

[0033] In other words, features of the measured H2 concentration and / or features of the measured H2O concentration can be used as input signals or input data for a trained machine learning method. Properties or features of the measured H2 concentration and / or the measured H2O concentration can be extracted, that is, the measured H2 concentration and / or the measured H2O concentration can be preprocessed to derive one or more features therefrom.

[0034] This means that a trained machine learning method can determine the H 2 concentration, N 2 concentration, steam concentration and / or water amount on the anode side of the fuel cell stack based on pre-processed measurements of H 2 concentration and / or H 2 O concentration.

[0035] Examples of features are peak height, area under the peak, and peak slope.

[0036] The H 2 concentration, N 2 concentration, steam concentration and / or water content on the anode side can thus be determined more reliably and more precisely, which in turn allows for more accurate and better optimization of the purge and bleed strategies.

[0037] Advantageously, the method further comprises the following steps:

[0038] - the purge strategy to be performed is derived from the determined H 2 concentration, the determined N 2 concentration, the determined steam concentration and / or the determined amount of water on the anode side of the fuel cell stack and / or

[0039] Or at least one parameter of the relief strategy.

[0040] In other words, the output signal of the trained machine learning method can be post-processed. From the output signal of the trained machine learning method, the executed purge strategy and bleed strategy or parameters of the purge process or bleed process can be extracted.

[0041] Advantageously, the method is applied at predefined time intervals for different power levels or operating parameters.

[0042] Advantageously, the H 2 concentration, N 2 concentration, steam concentration and / or water amount on the anode side of the fuel cell stack is determined using the trained machine learning method based on other system data, in particular based on other measured system data.

[0043] In other words, the machine learning method can be configured so that it can receive additional inputs or input signals. The input signals of the machine learning method can also be supplemented with other system data or signals, if necessary. For example, additional sensors can be provided that measure other system data and transmit them to the machine learning method. For example, an air mass flow sensor can be provided.

[0044] A second aspect of the invention relates to a method for training a machine learning method for a method for controlling a fuel cell system, as described above and below, having the following steps:

[0045] - generating a training data set with the aid of a mass spectrometer, a water quantity meter and an H2 sensor and / or an H2O sensor for the training fuel cell system,

[0046] - Create the input dataset by normalizing and post-processing the training dataset, and

[0047] - Training a machine learning method with an input data set, in particular a dynamic Gaussian process model.

[0048] Although the sensor used is referred to as an H2 sensor, it can be a sensor capable of detecting or measuring not only H2 concentration but also H2O concentration. The step of generating a training dataset can include performing one or more measurements using the H2 sensor, a water meter, and a mass spectrometer. The one or more measurements can serve as a knowledge base for training a machine learning method. The water meter can detect water quantity, that is, the amount of liquid water, and optionally also vapor concentration.

[0049] The training data set can be supplemented, if necessary, by other relevant data, such as the mass flow at the air inlet, the stack current or the anode outlet pressure, as well as by features or corrections extracted from these data.

[0050] The generated training dataset or generated data can then be processed to conform to the requirements of the machine learning method. In other words, the generated training dataset can be processed or manipulated so that the machine learning method can receive the processed or manipulated training dataset as an input dataset. Such processing or manipulation can include normalization and post-processing.

[0051] Normalization can be performed based on the mean and / or variance of the training dataset.

[0052] Advantageously, the method further comprises the following steps:

[0053] -Validate a trained machine learning method by using a subset of the input dataset.

[0054] Advantageously, the method for training a machine learning method further comprises the following steps:

[0055] - Testing the trained machine learning method to ultimately validate the machine learning method using validation datasets generated with the help of H2 sensors, water flow meters, and mass spectrometers, and

[0056] - determining the H 2 concentration, N 2 concentration, steam concentration and / or water content in the anode side of the fuel cell stack of the training fuel cell system by means of a tested, trained machine learning method based on the validation data set,

[0057] The determined H2 concentration, N2 concentration, steam concentration, and / or water amount on the anode side of the fuel cell stack of the training fuel cell system is compared with the H2 concentration and / or N2 concentration on the anode side of the fuel cell stack measured by the mass spectrometer. Additionally, the determined steam concentration and / or water amount on the anode side of the fuel cell stack of the training fuel cell system can be compared with the steam concentration and / or water amount on the anode side of the fuel cell stack measured by a water amount meter.

[0058] Although the sensor used is referred to as an H 2 sensor, it may be a sensor capable of detecting or measuring not only the H 2 concentration but also the H 2 O concentration.

[0059] Such a method for training a machine learning method can advantageously increase the accuracy and reliability with which the H 2 concentration, N 2 concentration, steam concentration, and / or water amount on the anode side can be determined. This allows for improved control of the fuel cell system and more precise and improved optimization of the purge and / or bleed strategies.

[0060] In the context of the present invention, the term "dataset," which may be a validation dataset, an input dataset, or a training dataset, refers to a collection of data pairs or data sets that include the corresponding input values ​​and output values ​​of a method, in particular a machine learning method. For example, a validation dataset may include not only the values ​​of the H2 signal and / or the H2O signal of the measuring line, or the H2 concentration and / or the H2O concentration, but also the corresponding values ​​of the mass spectrometer (H2 concentration and / or N2 concentration). Additionally, a validation dataset may include values ​​from a water meter.

[0061] All advantages described in detail with respect to the method according to the first aspect of the invention for optimizing a purge strategy for a fuel cell system also apply to the method according to the second aspect of the invention for training a machine learning method for a method for controlling a fuel cell system.

[0062] A third aspect of the present invention relates to a fuel cell system, in particular a PEM fuel cell system. The fuel cell system comprises a fuel cell stack with an anode side and a cathode side, an exhaust line, a fuel line with a recirculation loop, and at least one valve line connected to the recirculation loop. The at least one valve line and the exhaust line merge into a measuring line. The fuel cell system also comprises an H2 sensor and / or an H2O sensor arranged on the measuring line, and a control unit. The control unit is configured to perform the method as described above and below. The fuel cell system may also comprise a valve, which may be implemented as a purge and drain valve. Alternatively, the fuel cell system may comprise two valves, wherein one valve is a purge valve and the second valve is a drain valve. The fuel cell system may further comprise a water separator. In this case, the drain valve is preferably connected upstream of the water separator.

[0063] The advantages described with respect to the method for controlling a fuel cell system also apply to the fuel cell system according to the present invention. Therefore, such a fuel cell system is durable and can effectively reduce costs.

[0064] The same advantages as those described in the methods according to the first and second aspects of the present invention also apply to the fuel cell system according to the third aspect of the present invention.

[0065] A fourth aspect of the invention relates to a computer program product comprising instructions which, when executed by a control unit, cause the control unit to carry out the method for controlling a fuel cell system as described above and below.

[0066] A fifth and final aspect of the invention relates to a computer-readable medium on which the computer program product as described above is stored.

[0067] The advantages described in detail for the methods according to the first and second aspects of the present invention also apply accordingly to the computer program product according to the fourth aspect of the present invention and the computer-readable medium according to the fifth aspect of the present invention. All disclosures described above and below for one aspect of the present disclosure also apply to all other aspects of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Hereinafter, embodiments of the present invention will be explained with reference to the accompanying drawings.

[0069] Figure 1A fuel cell system according to an embodiment is schematically shown.

[0070] Figure 2 A schematic flow chart of a method according to an embodiment is shown.

[0071] Figure 3 A flow chart showing a method according to an embodiment is shown,

[0072] Figure 4 A fuel cell system according to an embodiment is schematically shown, and

[0073] Figure 5 The measurement of the purge process is shown schematically.

[0074] Similar, similarly acting, identical or identically acting elements are provided with similar or identical reference symbols in the figures.The figures are merely schematic and not drawn to scale. DETAILED DESCRIPTION

[0075] Figure 1 A fuel cell system 100 according to an embodiment is schematically shown. The fuel cell system 100 has at least one fuel cell stack 10 with an anode side 11 and a cathode side 12. The fuel cell system 100 also has an air path via which air from the surrounding environment can be supplied to the cathode side 12. An air compressor and / or a three-way valve 29 are arranged in the air path, which perform compression or suction depending on the respective operating conditions of the fuel cell stack 10. The three-way valve 29 makes it possible for air to bypass the cell stack when air is not required. In addition, a compressor 30 can be installed upstream of the three-way valve 29. Other components, such as filters and / or heat exchangers and / or valves, can also be provided within the air path. Oxygen-containing air can be supplied to the fuel cell stack 10 via the air path.

[0076] The fuel cell system 100 further comprises an exhaust gas line 13, a fuel line having a recirculation circuit 14, and a valve line 15 connected to the recirculation circuit 14. The valve line 15 and the exhaust gas line 13 merge into a measuring line 16.

[0077] A high pressure tank 28 and a shut-off valve 26 may be located in the inlet of the fuel line. Other components, such as a jet pump 24 or a blower 22, may be arranged in the fuel line to supply the anode side 11 of the fuel cell system 100 with fuel.

[0078] The valve line 15 has a valve 18. The valve 18 can be implemented as a purge and relief valve. The valve line 15 is arranged between the recirculation circuit 14 and the exhaust line 13 so that the gas mixture can flow from the recirculation circuit 14 into the measuring line 16. An H2 sensor 19 is arranged on the measuring line 16, which is configured to detect the H2 concentration and / or H2O concentration in the measuring line 16. A water meter (not shown here) can also be arranged on the measuring line 16. The water meter can detect the water content (i.e., the amount of liquid water) and / or the steam concentration. It should be noted that the H2 sensor and the water sensor can also be implemented as a single combined sensor.

[0079] Figure 1 The fuel cell system 100 is shown without the mass spectrometer 17. The training fuel cell system 100 may essentially have the same Figure 1 The same components or elements as the fuel cell system 100. Figure 1 Unlike the fuel cell system 100 of the training fuel cell system 100, the training fuel cell system 100 also has a mass spectrometer 17 and a water quantity meter (not shown here) arranged in the recirculation loop 14. The fuel cell system 100 that is not the training fuel cell system 100 preferably has neither a mass spectrometer 17 nor a water quantity meter in the recirculation loop.

[0080] Figure 1 The components and / or elements of the fuel cell system 100 can be controlled via a control unit (not shown here). In particular, the valve 18 used as a purge valve and as a drain valve can be controlled via such a control element.

[0081] Figure 2 A method for controlling a fuel cell system 100 according to an embodiment is shown, in particular as Figure 1 Schematic flow chart of a method for a fuel cell system 100 is shown in FIG. The method can be summarized into three method steps, wherein the third method step can be S3 or S4.

[0082] In a first step S1, measurement data are detected. In particular, the H2 concentration and / or H2O concentration is measured on the measuring line 16. Based on such measurements, corresponding data can be generated. Figure 2 The dashed arrow 31 represents the transmission of the measured H2 concentration to the trained machine learning method 20, which is symbolically represented by means of a box. Figure 2 The dashed arrow 32 represents the transmission of the measured H2O concentration to the trained machine learning method 20, which is symbolically represented by a box. Such transmission can be performed by means of a signal. Figure 2The dashed arrow 34 represents the optional transmission of other system data or system signals, such as air mass flow, to the trained machine learning method 20. Other inputs or input signals may also be transmitted to the trained machine learning method 20. The number of signals that the trained machine learning method 20 may receive as input is unlimited.

[0083] In a further step S2, the trained machine learning method 20 determines the H 2 concentration, N 2 concentration, steam concentration and / or water amount on the anode side 11 of the fuel cell stack 10, wherein the trained machine learning method 20 is based on the H 2 concentration and / or H 2 O concentration measured in the measuring line 16 for this purpose. In other words, the trained machine learning method 20 receives the measured H 2 concentration and / or H 2 O concentration as input signals 31, 32 and can provide the H 2 concentration, N 2 concentration, steam concentration and / or water amount on the anode side 11 as output signal 33 or output.

[0084] In a third step S3, the purge duration and / or the purge interval, as well as the bleed duration and / or the bleed interval, are adjusted or determined based on the determined H 2 concentration and / or N 2 concentration. In other words, in the third step S3, the purge strategy and the bleed strategy are determined and, if necessary, optimized by determining an optimized purge duration and / or an optimized purge interval, as well as an optimized bleed duration and / or an optimized bleed interval for the valve 18.

[0085] Instead of step S3 , only the discharge duration and / or the discharge interval may be determined in the third step S4 .

[0086] Figure 3 A flow chart of a method for training a machine learning method 20 according to one embodiment is shown. In a first step T1, a training data set is generated using the mass spectrometer 17, the water meter, and the H2 sensor 19 of the training fuel cell system 100. In other words, measurements are performed as a basis for the training data set. Such a data set can be used as an input data set. The training data set can include measurement data from the H2 sensor 19, which are provided with labels from the mass spectrometer 17 and from the water meter. This knowledge base can thus encompass the widest possible range of system behavior. It has proven particularly advantageous to take into account at least one of the following points when measuring with the H2 sensor 19:

[0087] The mass spectrometer 17 and the water quantity meter should remain switched on during the measurement, since they provide a label for each measured H 2 concentration and / or H 2 O concentration, ie a target value for the output of the machine learning method 20 .

[0088] The measurements of the H2 sensor 19 should cover the largest possible relevant load range, so that the machine learning method 20 trained thereby can have a wide range of applicability.

[0089] The H2 sensor 19 should display a peak during each purge and / or bleed process. Displaying a zero value over multiple purge and / or bleed processes should be avoided. This can occur particularly at high current intensities, as a high air mass flow is fed through the cathode, which significantly dilutes the hydrogen mass flow in the exhaust gas.

[0090] The measurements of the H2 sensor 19 should include different system dynamics. To this end, transient measurements can be performed at different current intensities. For example, a step-shaped current signal between 50, 100, 120, and / or 237 A can be run. Each time the current intensity is switched, transient processes can be generated in the mass spectrometer 17 and / or the water meter. By subsequently maintaining the current level for a certain period of time, both dynamic and static system behavior can be incorporated into the training data.

[0091] As a result, the machine learning method 20 trained in this manner can more accurately and reliably determine the H2 concentration, N2 concentration, steam concentration, and / or water content. If no new operating strategy is implemented and the aging state of the fuel cell stack has not significantly changed, the measurements used to generate the training data set may only need to be performed once. If these changes occur, the measurements may need to be repeated to retrain the machine learning method 20. Such repetitions can be performed, for example, in the workshop during scheduled maintenance or as a preventive measure in the event of an error message from, for example, the H2 sensor 19.

[0092] In a further step T2, the training dataset is normalized and post-processed to create an input dataset. In other words, the training dataset is prepared for the machine learning method 20 to be trained. To this end, the training dataset may be normalized and outliers may be removed. Filtering of the training dataset may be omitted. However, filtering of the training dataset may be performed if desired.

[0093] In a further step T3, the machine learning method 20 is trained. The machine learning method 20 can be trained in particular based on a Gaussian process model.

[0094] In the next step T4, the trained machine learning method 20 is validated. To this end, cross-validation can be performed based on the training dataset, in particular to avoid overfitting. The machine learning method 20 can then be validated based on other datasets with different system dynamics. The input data of the validation dataset should also be normalized, for example, based on the mean and variance of the training dataset.

[0095] If a large validation error exists, it may be advantageous to verify whether the measurements performed during the step of generating the training dataset T1 meet certain requirements and / or whether changes in the operating strategy of the fuel cell system 100 have been implemented. If these conditions can be excluded, the problem can be resolved by supplementing the input dataset with one or more features, if necessary. Otherwise, steps T1 and / or T5 can be repeated.

[0096] In a final, further step T6, the trained machine learning method 20 can be tested. Thus, a final validation of the trained machine learning method 20 can be performed. To this end, the trained machine learning method 20 is tested with additional measurement data that is independent of the training data and / or validation data. A test data set can thus be generated by the H2 sensor 19 and by the water meter, which differs both from the training data set and from the validation data set.

[0097] Figure 4 A fuel cell system 100 according to an embodiment is schematically shown. Unless otherwise specified, Figure 4 The fuel cell system has Figure 1 The same elements and / or components as the fuel cell system of Figure 1 The fuel cell system is different. Figure 4 The valve line 15 of the fuel cell system is divided into two lines. A purge line 15.1 and a drain line 15.2 are provided, which are connected in parallel to each other. A purge valve 25 is arranged on the purge line 15.1, and a drain valve 21 is arranged on the drain line 15.2. The drain valve 21 is basically connected upstream of the water separator (not shown here). In the training fuel cell system, a mass spectrometer 17 and a water quantity meter can be arranged in the recirculation loop 14. In addition, a training drain line 23 can be provided in the training fuel cell system. It is conceivable to provide a water quantity meter on the training drain line 23, which can detect not only the amount of water in the training drain line but also the steam concentration therein.

[0098] Figure 5 The measurement of the purge process is shown schematically. Time t is plotted on the horizontal axis. The vertical axis is dimensionless and is intended only to schematically represent the changes in the corresponding variables. The solid line 44 represents the state of the purge valve 25, wherein the vertical offset corresponds approximately to the opening of the purge valve 25. The dashed line 40 represents the measured H2 concentration in the measuring line 16. The dashed line 42 represents the determined H2 concentration on the anode side 11. The slope of the dashed line is greatest in the time segment W. The relatively steep slope of the dashed line, that is, the measured H2 concentration, allows deduction of the steam concentration and the nitrogen concentration. A relatively flat slope can, for example, indicate a relatively high nitrogen content.

[0099] It should be noted that the terms "comprising" and "having" do not exclude other elements, and the indefinite article "a" or "an" does not exclude a plurality. It should also be noted that features and steps described with reference to one of the above embodiments can also be used in combination with other features and steps of other embodiments described above. Reference signs in the claims should not be construed as limiting.

Claims

1. A method for controlling a fuel cell system (100), wherein: The fuel cell system (100) comprises a fuel cell stack (10) with an anode side (11) and a cathode side (12), an exhaust line (13), a fuel line with a recirculation circuit (14), and at least one valve line (15) connected to the recirculation circuit (14), wherein the at least one valve line (15) and the exhaust line (13) merge into a measuring line (16) and the at least one valve line (15) comprises a valve (18), The method comprises the following steps: - (S1) measuring the H2 concentration and / or H2O concentration in the measuring line (16), (S2) determining, based on the measured H2 concentration and / or H2O concentration, an H2 concentration, an N2 concentration, a steam concentration and / or a water amount on the anode side (11) of the fuel cell stack (10) by means of a trained machine learning method (20), and (S3) adjusting the purge duration and / or purge interval based on the determined H2 concentration and / or N2 concentration on the anode side (11) of the fuel cell stack (10), and adjusting the bleed duration and / or bleed interval based on the determined steam concentration and / or water amount on the anode side (11) of the fuel cell stack (10), or - (S4) adjusting a bleeding duration and / or a bleeding interval based on the determined steam concentration and / or water amount on the anode side (11) of the fuel cell stack.

2. The method according to claim 1, It is characterized by: The valve (18) of the at least one valve line (15) is a purge and bleed valve.

3. The method according to claim 1 or 2, It is characterized by: The fuel cell system (100) has two valve lines (15), wherein a first valve line (15.1) has a purge valve and a second valve line (15.2) has a drain valve (21).

4. The method according to at least one of the preceding claims, It is characterized by: The trained machine learning method (20) is a Gaussian process model, in particular a Gaussian process model with a NARX structure.

5. The method according to at least one of the preceding claims, It is characterized by: The trained machine learning method (20) is trained with the aid of a training fuel cell system, wherein the training fuel cell system comprises a fuel cell stack with an anode side and a cathode side, an exhaust line, a fuel line with a recirculation loop, at least one valve line (15) connected to the recirculation loop, a mass spectrometer (17) and a water quantity meter, wherein the mass spectrometer (17) is configured to detect an H2 concentration and / or an N2 concentration on the anode side of the training fuel cell system and the water quantity meter is configured to detect an amount of water on the anode side of the training fuel cell system.

6. The method according to at least one of the preceding claims, It is characterized by: The method further comprises the following steps: - deriving at least one characteristic of the respective measurement from the measured H 2 concentration and / or from the measured H 2 O concentration, and - Based on the at least one characteristic derived from the corresponding measurement, determining the H2 concentration, N2 concentration, steam concentration and / or water amount on the anode side (11) of the fuel cell stack (10) with the aid of a trained machine learning method (20).

7. The method according to at least one of the preceding claims, It is characterized by: The method further comprises the following steps: - deriving at least one parameter of an executed purge strategy and / or bleeding strategy from a determined H2 concentration, a determined N2 concentration, a determined steam concentration and / or a determined amount of water on the anode side (11) of the fuel cell stack.

8. The method according to at least one of the preceding claims, It is characterized by: The method is applied at predefined time intervals for different power levels or operating parameters.

9. The method according to at least one of the preceding claims, It is characterized by: The step of determining the H2 concentration, N2 concentration, steam concentration and / or water amount on the anode side (11) of the fuel cell stack (10) is performed with the aid of the trained machine learning method (20) based on other system data, in particular based on other measured system data.

10. Method for training a machine learning method (20) for a method for controlling a fuel system (100) according to any one of claims 1 to 9, The following steps are involved: (T1) generating a training data set with the aid of a mass spectrometer, a water quantity meter and an H2 sensor and / or an H2O sensor (19) for training a fuel cell system, - (T2) creating an input dataset by normalizing and post-processing the training dataset, and -(T3) training the machine learning method by means of the input data set, in particular a dynamic Gaussian process model.

11. The method according to claim 10, further comprising the steps of: -(T4) validating the trained machine learning method using a subset of the input dataset.

12. The method according to claim 10 or 11, further comprising the steps of: - (T6) testing the trained machine learning method (20) to ultimately validate the machine learning method (20) by generating a validation data set with the aid of the H2 sensor, the water meter and the mass spectrometer, and - determining, based on the validation data set, the H 2 concentration, the N 2 concentration, the steam concentration and / or the amount of water on the anode side of the fuel cell stack of the training fuel cell system by means of the tested, trained machine learning method, in, The determined H2 concentration and / or N2 concentration on the anode side of the fuel cell stack of the training fuel cell system is compared with the H2 concentration and / or N2 concentration on the anode side of the fuel cell stack measured by a mass spectrometer (17).

13. A fuel cell system (100), in particular a PEM fuel cell system, comprising: A fuel cell stack (10) with an anode side (11) and a cathode side (12), an exhaust line (13), a fuel line with a recirculation circuit (14) and at least one valve line (15) connected to the recirculation circuit, wherein: The at least one valve line (15) and the exhaust line (13) merge into a measuring line (16), and the fuel cell system (100) further comprises an H2 sensor and / or an H2O sensor and a control unit arranged on the measuring line (16), wherein the control unit is configured to execute the method according to any one of claims 1 to 9. 14 . A computer program product comprising instructions which, when executed by a control unit, cause the control unit to carry out the method according to claim 1 .

15. A computer-readable medium having stored thereon the computer program product according to claim 14.

Images